What survives when a PDF becomes a Word file

Understanding why tells you which tool to pick and how much to expect from it.

A PDF is a printout, not a document

What sits inside a PDF is a long list of instructions saying print this character at x 72, y 680. Nowhere in the file is it recorded where a paragraph ended, or which two lines belonged to the same row of a table.

That information is thrown away the moment a document is exported to PDF, in the same way a sheet of paper has no need of it. So turning one back is reconstruction rather than conversion: the original shape is inferred from where the characters sit.

That single sentence explains the rest. A table comes out crooked not because the converter misread the table, but because there was no table to read and one had to be inferred from position alone.

What can be inferred and what was never there

Separating these two tells you what can improve and what cannot.

• Inferable. Paragraph breaks, from line spacing and where lines stop. Tables, from whether cells line up down the page. Headings, from type larger than the body. Better algorithms make these more accurate.
• Directly extractable. The characters themselves, and any images or stamps printed on the page. These are real objects in the file and need no guessing.
• Never recorded. Which Word style the original used, where a table had merged cells, how many points of spacing sat between paragraphs. No tool can produce these.

Be wary of converters that reproduce the layout exactly

Some tools advertise output that matches the original pixel for pixel. Open the result and it genuinely does. Then you edit one line and the document collapses.

Achieving that means putting the text into hundreds of floating text boxes pinned to fixed coordinates rather than into paragraphs. A document built that way looks like the original but cannot be worked in. Delete a character and only that one box reflows while everything around it stays put.

Conversion exists so that you can edit. Output you cannot edit has lost the point. If looking identical is the goal, there was no reason to convert in the first place.

Why tables really come out wrong

Tables in PDFs very often have no ruling lines at all. Where lines do exist they are drawn on a different layer from the text, invisible to anything pulling the words out. That is why approaches built on finding the lines fail so often.

So position is what gets read: cells whose left edges stand in the same place across several lines are treated as a table. The method is weakest on empty cells. An empty cell leaves no trace in the file, so that row simply appears to have one cell fewer and everything after it shifts along by one.

Merged cells behave the same way. Your eye sees two cells joined; the file contains one piece of text sitting slightly further left than its neighbours.

No tool can convert a scanned PDF

A PDF produced by an office scanner is a single photograph per page. Your eyes see letters, the file contains none. This is not a limitation of the tool but an absence in the source.

Optical character recognition can manufacture text from the picture. But that is recognition rather than extraction, and it mixes in errors like reading a zero as the letter O. In a contract or a submitted form that kind of error is expensive. Asking the issuer for the original file is usually faster.

In short

The question to ask of a converter is not how closely it matches, but whether you can edit what comes out. And when the result is imperfect, being able to tell a limit of the tool from a limit of the format saves you from swapping tools for nothing.